Automox MCP Server

Automox MCP Server

The official MCP server for Automox. Talk to your Automox console using natural language — this MCP server connects AI assistants like Claude to your Automox environment so you can manage devices, check compliance, run policies, and more, just by asking.

You:   "Are we ready for Patch Tuesday?"
Claude: Here's your readiness summary — 3 devices need patches,
        2 approvals are pending, and your patch policies run tonight at 2 AM...

[!IMPORTANT] For bug reports or feature requests, use help.automox.com or your typical escalation paths.

[!CAUTION] AI assistants can make mistakes. Responses produced by the MCP server may be incorrect or incomplete. If you see this happening consistently, please let us know.

Table of Contents

• What's New in 3.0
• Self-Hosted vs. Hosted
• Quick Start
• What Can I Ask?
• Configuration
• Security
• Privacy Policy
• Alternative Installation
• Updating
• Migrating to the Hosted Server
• Troubleshooting
• Frequently Asked Questions
• Development
• Versioning
• License
• Support

What's New in 3.0

Automox MCP Server 3.0 adds a centrally hosted option to replace the legacy self-hosted solution. The hosted server supports all the functionality you're already using, plus one new capability.

• Hosted server: Automox now runs and maintains a version of this same server for you. Nothing to install, connect your AI client to https://console.automox.com/api/mcp with your existing API key. See Self-Hosted vs. Hosted below.
• Policy Catalog templates (hosted only): The hosted service can search Automox's library of best-practice policy templates and create a policy directly from one, so you have a starting point without building a policy from scratch. This is separate from asking about the policies already deployed in your org, this is Automox's own recommended template library.
• No capability loss: The hosted service exposes the same tool coverage you already have through this repository.
• No forced migration: This self-hosted server and the Claude Desktop extension keep working exactly as they do today. Moving to the hosted service is currently optional (but recommended), see Migrating to the Hosted Server.
• Auth is unchanged for now: Both versions use the same Automox API key. SSO is planned for a future release.

Self-Hosted vs. Hosted

Both run the same open-source server code. The difference is who installs and runs it, and one capability that's currently hosted-only.

For self-hosted setup, see Quick Start below. For the hosted server, see Migrating to the Hosted Server, which also covers connecting fresh with no prior install.

Quick Start

1. Get your Automox credentials

You need three values from the Automox Console:

Automox has two API key types, and the difference matters here:

Symptom: 403 on the search tools while reads work everywhere else usually means the key, not your permissions — switch to an org-scoped key for the target org. API Key and Account UUID are always required. Org ID is recommended but optional — some tools that don't require org context will work without it.

2. Create a .env file

AUTOMOX_API_KEY=your-api-key
AUTOMOX_ACCOUNT_UUID=your-account-uuid
AUTOMOX_ORG_ID=your-org-id

3. Connect to your AI assistant

Claude Desktop (recommended) — one-click MCPB install:

New: Automox MCP is now also listed directly in Claude's Connectors Directory. Open Claude Desktop, go to Settings > Connectors, and search "Automox" to connect without leaving the app. The manual steps below still work too, nothing about them has changed.

1. Download the latest automox-mcp-<version>.mcpb from the GitHub Releases page.
2. Open Claude Desktop → Settings → Extensions.
3. Drag the .mcpb file into the Extensions window.
4. Paste your API key, Account UUID, and (optionally) Org ID into the prompts.

No .env file, no terminal — credentials are stored in Claude Desktop's secure config. The bundle pulls the matching automox-mcp release from PyPI on first run.

Claude Code (CLI):

claude mcp add automox-mcp uvx -- --env-file /path/to/.env automox-mcp

Cursor / any other MCP client — add to your MCP config:

{
  "mcpServers": {
    "automox-mcp": {
      "command": "uvx",
      "args": ["--env-file", "/path/to/.env", "automox-mcp"]
    }
  }
}

That's it. Start asking questions.

What Can I Ask?

The server exposes 130+ tools across devices, policies, patches, groups, webhooks, worklets, vulnerability sync, maintenance windows, and more. You don't need to know the tool names — just describe what you want:

For the full list of tools, parameters, and MCP resources, see the Tool Reference.

Tip: You can also ask the server itself — the discover_capabilities tool returns all available tools organized by domain.

Configuration

Environment Variables

Applies only to self-hosted 2.x servers. Not applicable to the hosted 3.0+ server.

Read-Only Mode

AUTOMOX_MCP_READ_ONLY=true

Disables all write operations. Only read-only tools are registered (85 of 133). Useful for auditing and monitoring.

Modular Loading

Load only the tool modules you need:

AUTOMOX_MCP_MODULES=devices,policies

Available modules: audit, audit_v2, devices, device_search, policies, policy_history, users, groups, events, reports, packages, webhooks, worklets, data_extracts, vuln_sync, compound, policy_windows

Both settings can be combined:

AUTOMOX_MCP_READ_ONLY=true
AUTOMOX_MCP_MODULES=devices,policies

HTTP Transport

For non-stdio deployments:

uvx --env-file .env automox-mcp --transport http --host 127.0.0.1 --port 8000

Endpoint Authentication

When deploying over HTTP or SSE, you can require authentication on the MCP endpoint (separate from the Automox API key). Two strategies are supported:

Static API keys (simple):

automox-mcp --generate-key                         # generate a key
export AUTOMOX_MCP_API_KEYS="amx_mcp_a1b2c3..."    # or use a key file

OAuth 2.1 / JWT (enterprise IdP integration):

export AUTOMOX_MCP_OAUTH_ISSUER="https://auth.example.com/realms/main"
export AUTOMOX_MCP_OAUTH_AUDIENCE="https://mcp.example.com"
export AUTOMOX_MCP_OAUTH_SERVER_URL="https://mcp.example.com"  # enables RFC 9728 metadata

Clients must include Authorization: Bearer <token> on every request. Unauthenticated requests receive 401 Unauthorized with proper WWW-Authenticate headers. No effect on stdio transport.

Security

The Automox MCP server is designed for enterprise deployment with defense-in-depth security controls.

Highlights:

• Read-only mode (AUTOMOX_MCP_READ_ONLY) disables all 48 write tools
• Module filtering (AUTOMOX_MCP_MODULES) for least-privilege tool loading
• Correlation IDs on every tool call, forwarded to Automox API as X-Correlation-ID
• Rate limiting (30 calls/60s) with token budget estimation and auto-truncation
• API key isolation — stored as private attribute with per-request auth injection (no header storage)
• Generic error responses — no internal paths, connection strings, or API keys in error output
• Prompt injection mitigation — API response sanitization with Unicode normalization, homoglyph defense, HTML tag/script stripping, and reference-style markdown stripping
• Webhook secret handling — secrets stripped from idempotency cache after creation
• Structured JSON logging (AUTOMOX_MCP_LOG_FORMAT=json) for SIEM integration
• Tool name prefixing (AUTOMOX_MCP_TOOL_PREFIX) to prevent cross-server collisions
• Sigstore-signed releases with CycloneDX SBOM
• SSRF prevention — webhook URLs validated against private/loopback IPs and cloud metadata endpoints
• MCP endpoint authentication — static API keys or OAuth 2.1/JWT with audience binding and RFC 9728 Protected Resource Metadata
• DNS rebinding protection — Origin and Host header validation on all HTTP/SSE connections per the MCP transport spec
• Security response headers — X-Content-Type-Options, X-Frame-Options, CSP, Cache-Control: no-store, Strict-Transport-Security on all HTTP responses
• Authentication rate limiting — blocks IPs after repeated auth failures to mitigate brute-force attacks
• Remote bind protection — non-loopback HTTP/SSE binding requires explicit --allow-remote-bind opt-in
• MCP Tool Annotations on all 130+ tools — readOnlyHint, destructiveHint, idempotentHint, and openWorldHint per the MCP Protocol specification, enabling client-side confirmation dialogs and safety guardrails
• Interactive MCP Apps (io.modelcontextprotocol/ui) — inline review/approval surfaces for consequential flows: compliance triage, patch approval, policy blast-radius, remediation apply, and RBAC access certification. Apps-capable hosts render them inline; other hosts degrade gracefully to the structured tool output. Write-flow Apps drive the existing gated tools through the host's confirmation — no new tools, no new gates — and ship under the host's deny-all CSP (self-contained, no external/CDN loads)
• 61 security hardening items (V-001 through V-182, S-001 through S-006) documented in CHANGELOG and SECURITY.md

Capability model. The server wraps 100% of the published Automox Console API and Webhooks API, with a single deliberate exception — secret-exposing endpoints are never wrapped (API-key decrypt, password-setting). Every destructive operation is either ask-first (host confirmation) or gated behind a default-off env flag. Concretely, three categorical rules:

• Secrets are never handled — the server never returns secret material and never lets the model set it. Credentials enter only via environment/config; decrypt endpoints are not wrapped, password-setting is excluded, and secret fields are redacted from every projection. This is the only intentional omission.
• Destructive operations are two-tier. Single-target, recoverable actions are ask-first (destructiveHint: true, surfaced as a host confirmation dialog, disabled entirely by read-only mode). Operations where per-call confirmation can't protect you — fleet-scale, self-lockout, or arbitrary model-authored code execution — are gated behind explicit, default-off env flags (AUTOMOX_MCP_ALLOW_APPLY_REMEDIATION_ACTIONS, AUTOMOX_MCP_ALLOW_SPLASHTOP_BULK_INSTALL_UNINSTALL, AUTOMOX_MCP_ALLOW_DELETE_DEVICE). Device deletion is gated, not omitted.

The full coverage map, the gating principle, and every intentional omission are documented in API Coverage & Intentional Omissions.

For vulnerability reporting and the full threat model, see SECURITY.md. For deployment hardening (containers, Kubernetes, MCP gateways, TLS, authentication), see the Deployment Security Guide. Security posture is benchmarked against the Wiz MCP Security Best Practices cheat sheet.

Note: For network-accessible deployments, enable endpoint authentication (static keys via AUTOMOX_MCP_API_KEYS or JWT